Claude Fable 5.2: specs and benchmark results

Sources checked 11 November 2026.

Anthropic released Claude Fable 5.2 on 9 November 2026. On Anthropic’s published benchmarks, it scores above Fable 5.1 on every listed measure and above Opus 5.5 on most. Claude Fable 5.2 is available on Autohive now.

A token is a small chunk of text, often part of a word. Models measure what they read (input) and write (output) in tokens. Effort is how hard the model thinks before answering. Higher effort can improve an answer.

Claude Fable 5.2 specs at a glance

Specs are from Anthropic’s models overview.

What changed from Fable 5.1

Anthropic says the model writes tokens up to 30% faster. Its fast mode writes up to 2.5 times faster. The Remote Labor Index score more than doubled, from 17.9% to 41.3%. AutomationBench rose 20 percentage points.

Anthropic’s Fable 5.2 announcement describes an earlier build of the model migrating a service from Python to Rust. The migrated service started 26 times faster. That’s Anthropic’s demo, not a gain Autohive has tested.

Claude Fable 5.2 benchmarks

These scores come from Anthropic’s Fable page, not Autohive testing. Safeguards were on. Each model ran at its highest available effort unless marked below. Effort wasn’t matched across models, so the gaps need care. Scores are percentages except Elo.

Opus 5.5 has the higher AA-Briefcase Elo, 1800 to 1785. On AutomationBench, Gemini 4 Carbon scored 53.9%, ahead of Fable 5.2’s 51.4%. The OSWorld score gives partial credit in an offline setup. It isn’t the share of tasks fully completed.

The widest leads over Opus 5.5 in this table are Remote Labor Index (41.3% vs 21.3%) and Terminal-Bench-Science (74.6% vs 58.7%).

An independent vision test

Ultralytics’ auto-annotation test, published 10 November, covered 360 images from 36 datasets. Fable 5.2 placed second of 44 models overall. It ranked first on mAP50-95 (0.347), a stricter measure of object-detection accuracy.

It also refused all 10 baggage X-ray images, where Fable 5.1 refused one. Ultralytics tied this to filters around explosives. Eleven requests failed, including one timeout. These small samples show where safeguards can block legitimate work. They can’t predict refusals on your own images.

Safeguards

Fable 5.2 shares its underlying model with the restricted Claude Mythos 5.2. Fable adds safeguards for cybersecurity and biology. If your work covers security research, lab science or related topics, test real examples before relying on it.

Claude Fable 5.2 vs Opus 5.5

The published results make Fable 5.2 worth testing for long, multi-step work, coding and chart reading. Keep Opus 5.5 for routine work it already handles well.

Test both on the same tasks and compare accuracy and retries. Our guides to choosing AI models and routing tasks to different models explain the choices.

Claude Fable 5.2 in Autohive

Claude Fable 5.2 is available to select for your agents in Autohive. You choose a model per agent. Agents use the actions and integrations you’ve authorised, within an execution sandbox rather than with free desktop control.

There’s no manual effort slider. Fast mode and the full context and output limits aren’t confirmed for Fable 5.2 in Autohive.

FAQ

Can I use Claude Fable 5.2 in Autohive yet?

Yes. Claude Fable 5.2 is available on Autohive now. You can select it as the model for an agent. Before switching an agent you rely on, test Fable 5.2 on the same tasks and check the answers against the results you expect.

How much text can Claude Fable 5.2 handle?

Through Anthropic’s API, Fable 5.2 can read up to 1M tokens in one request and write up to 128K tokens in a single response. These are Anthropic’s provider limits. Autohive hasn’t confirmed its own context and output settings for Fable 5.2.

What’s the difference between Claude Fable and Claude Mythos?

They use the same underlying model. Fable is the generally available version, with added cybersecurity and biology safeguards. Mythos access is restricted. Anthropic publishes some Mythos-only results for cyber and biology tasks. Those don’t describe Fable’s behaviour, so they aren’t included in our benchmark table.